Accuracy to detection timing for assisting repetitive facilitation exercise system using MRCP and SVM
Miura et al. Robot. Biomim.
Accuracy to detection timing for assisting repetitive facilitation exercise system using MRCP and SVM
Satoshi Miura 0
Junichi Takazawa 2
Yo Kobayashi 1
Masakatsu G. Fujie 1
0 Faculty of Science and Engineering, Waseda University , 3-4-1, Okubo, Shinjuku-ku, 169-8555 Tokyo , Japan
1 Healthcare Robotics Institute, Future Robotics Organization, Waseda University , 3-4-1, Okubo, Shinjuku-ku, 169-8555 Tokyo , Japan
2 Graduate School of Science and Engineering, Waseda University , 3-4-1, Okubo, Shinjuku-ku, 169-8555 Tokyo , Japan
This paper presents a feasibility study of a brain-machine interface system to assist repetitive facilitation exercise. Repetitive facilitation exercise is an effective rehabilitation method for patients with hemiplegia. In repetitive facilitation exercise, a therapist stimulates the paralyzed part of the patient while motor commands run along the nerve pathway. However, successful repetitive facilitation exercise is difficult to achieve and even a skilled practitioner cannot detect when a motor command occurs in patient's brain. We proposed a brain-machine interface system for automatically detecting motor commands and stimulating the paralyzed part of a patient. To determine motor commands from patient electroencephalogram (EEG) data, we measured the movement-related cortical potential (MRCP) and constructed a support vector machine system. In this paper, we validated the prediction timing of the system at the highest accuracy by the system using EEG and MRCP. In the experiments, we measured the EEG when the participant bent their elbow when prompted to do so. We analyzed the EEG data using a cross-validation method. We found that the average accuracy was 72.9% and the highest at the prediction timing 280 ms. We conclude that 280 ms is the most suitable to predict the judgment that a patient intends to exercise or not.
Neurorehabilitation; Repetitive facilitation exercise; Brain-machine interface; Motor command detection
The number of cerebral stroke patients is increasing
worldwide. For example, in Japan, cerebral stroke patients
exceeded 2.8 million people in 2015 [
]. Patients often
suffer from aftereffects following a stroke, the most
frequent of which is hemiplegia. To recover motor function
following hemiplegia, patients must endure a long course
of difficult rehabilitation. Many studies have investigated
methods to shorten the recovery time through efficient
rehabilitation after hemiplegia [
Neurorehabilitation has been shown to be very efficient
]. Neurorehabilitation is a method to prompt the
recovery of the injured neural system. Repetitive
facilitation exercise is drawing attention as a particularly
effective rehabilitation. Kawahira demonstrated the efficacy of
repetitive facilitation exercise . Patients using
repetitive facilitation exercise can recover motor control three
times faster than those using usual therapy [
facilitation exercise also improves the paralysis part to
health compared with usual therapy [
Figure 1 shows the mechanism of the repetitive
facilitation exercise. The patient imagines moving a paralyzed
body part. Within the patient’s body, motor commands
travel from the brain, through the spinal cord, to the
paralyzed part. The therapist then stretches the paralyzed
part of the patient using physical or electrical
stimulation before the motor command reaches the spinal cord.
This stimulation excites the nerves in the spinal cord and
activates the path of the motor command. Because the
motor command can pass more easily through the nerve
pathway, the patient becomes likely to regain the ability
to move the paralyzed part unaided.
The most important point of repetitive facilitation
exercise is the stimulation timing. It is necessary for the
stimulation to occur before the motor command reaches
the spinal cord for success of the repetitive facilitation
exercise. However, it is difficult to successfully perform
repetitive facilitation exercise because even a skilled
therapist cannot detect the timing of the motor command.
The success of repetitive facilitation exercise is
dependent on the intuition and experience of the therapist.
Even an experienced therapist cannot achieve a 100%
success rate with the repetitive facilitation exercise
because therapists cannot detect patient motor
intentions. It has been reported that five of twelve patients
showed increased motor evoked potential after repetitive
facilitation exercise by therapists [
]. This shows that
the nervous systems of these patients underwent some
reconstruction. Although there are individual differences,
it is said that the conventional success rate is about five
of twelve, that is just 45%. To improve the higher success
rate than the target value 45%, therapists require a system
to assist repetitive facilitation exercise. An upper limb
reaching device has been proposed to reduce fatigue and
pain in the paralyzed arm and to decrease the burden on
the therapist [
]. The patient repeats inward and outward
movements to push the front and back buttons
alternately. When the patient pushes the button, the device
generates vibrations and electric stimulation to make it
easier to move the paralyzed arm. However, the device
cannot control the accurate stimulation timing because
the device cannot detect the motor command generated
in the brain.
There have been many studies of real time
detecting the motor command in brain. Most of these have
used electroencephalogram (EEG) because it has a
higher time resolution than other brain
measurement devices. EEG is simple to analyze in real time.
For example, Lucian reported that EEG can show the
steering timing of a driver during driving [
10, 11, 12
He measured and analyzed EEG data while
participants operated a driving simulator. Using this method,
the turn direction was detected 811 ms before
steering with an accuracy of 74.6%. In another study, Choi
showed that a brain–machine interface system using
EEG could be used to control a wheelchair [
system analyzed EEG data and moved forward or
turned left or right based on the measured EEG
signals within 125 ms.
These studies are useful for realizing real-time brain–
machine interfaces for healthy individuals. However,
these studies have not been adapted to the rehabilitation.
Our motivation is to develop a brain–machine interface
system to assist repetitive facilitation exercise. The
system detects motor intention in real time by measuring
the brain and stimulates the paralyzed body part before
the motor command reaches the spinal cord, as shown
in Fig. 2. The system overview and flow are shown in
Figs. 3 and 4. The system is consisted of the EEG
measurement device and FES generator, shown in Fig. 3. The
system detects motor commands from the patient’s EEG
data using support vector machine (SVM). It then
stimulates the paralyzed part by functional electric stimulation
(FES) before the motor command reaches the spinal cord.
The system provides stimulation at the same time the
SVM detects the motor command; thus, the stimulation
Fig. 2 Proposed system. The system measures the patient’s EEG
using the EEG amp and analyzes EEG data using the SVM. The system
actuates FES using the stimulation device. The system stimulates the
paralyzed part electrically
occurs after the motor command is produced, but before
it reaches the spine.
In this paper, we validate our proposed motor
command detecting method. The prediction accuracy
changes as to the timing before action. It is not clear that
the accuracy changes as to the prediction timing
changing. We clarify the appropriate prediction timing by the
accuracy of the system using SVM. We constructed an
SVM system and carried out experiments to clarify the
detection ratio of motor commands. In the experiment,
we collected EEG data when the participant bent their
elbow, shown in Fig. 5. To facilitate the timing, we
displayed a bar on a monitor to indicate to the participant
when to bend their elbow. The bar was displayed on the
monitor and shortened gradually until it disappeared.
At the same time the bar disappeared, the participant
had to bend their elbow. We analyzed EEG data using a
cross-validation method to clarify the detection ratio of
the motor command. We confirmed that the calculated
detection ratio was above the target value, verifying the
utility of the proposed system.
Three healthy participants (male, age 22–23, two
righthanded and one left-handed) were enrolled in the
experiment. We did not enroll paralyzed patients because the
aim of the present study was only to validate the
proposed method for detecting motor commands. Informed
consent was obtained from all participants. All
participants attested to having slept well the night before the
experiment to exclude the influence of sleep deprivation.
All participants did not intake the drugs such as caffeine,
alcohol, nicotine, and other medicines. The experiments
were approved by the Waseda University Institutional
Review Board (No. 2014-156).
We used an EEG (g. USBamp, gtec, USA) to measure
brain activity. We set up the device as shown in Table 1.
The sampling frequency was set to 256 Hz because the
real-time measurement using MRCP needs a high
temporal resolution. We used 17 analog input channels and
1 GND passive channel. On an EEG cap, 14 channels
were located to measure EEG based on the international
10–20 system, as shown in Fig. 6. The reference electrode
was put on the earlobe.
We used a three-degree-of-freedom accelerometer
(ACL300, Biometric Inc., USA) to detect the time when the
participant moved their arm. We affixed the accelerometer
to the participant’s wrist as shown in Fig. 6 and connected
it to an analog input–output board (AIO-163202FX-USB,
Contec Inc., USA) to get the analog input value. The
accelerometer used three channels, as shown in Table 1.
We conducted the experiment in a closed room to
minimize noise disturbances. During the experiment, the
participant did not talk and sat still. In addition, we asked the
participant to try to avoid swallowing saliva or blinking
hard. The temperature was 20 ± 15 °C, and the humidity
We put the EEG cap and the accelerometer on the
participant. The experimental procedure is shown in Fig. 8.
In the initial state, the participant relaxed with their right
forearm resting on the desk, palm up. One measurement
session consisted of rest, preparation and act periods
within 10 s. During the rest period, from 0 to 3 s, the
participant relaxed and looked at the monitor. During the
preparation period, from 3 to 6 s, a red bar was displayed
on the monitor and became smaller until disappearing at
6 s, as shown in Fig. 7. During the act period, after 6 s, at
the same time as the red bar disappeared, the participant
bent their arm at the elbow. The participant kept their
elbow bent for approximately 0.5 s and then set their arm
to the initial resting state. This experimental procedure
was performed 100 times by each participant.
The experimental condition is the prediction timing.
The prediction accuracy changes as to the timing before
Fig. 8 Classification by the support vector machine (SVM). The
separation plane divides the data into Class 1 and Class 2. The margin
is the distance between the separation plane and the closest data
point. SVM sets the separation plane to maximize the margin
training sample data. SVM decided the two outputs using
y = sign W T x − h
action. However, the timing is influenced by the
recognition delay. Even the participant bent his/her elbow as
soon as the disappearance of the bar, and there is actually
the delay because it takes time to recognize it. The
recognition delay is said to about 200 ms, but there is no clarity
of accurate delay. In this paper, the experimental
condition is the prediction timing around 200 ms before action
timing. The condition is 70, 140, 210, 280 and 350 ms
before action timing.
Movement‑related cortical potential (MRCP)
We focused on the functions of EEG. For
example, event-related potential is the electric fluctuation
detected from neurons following light or sound
]. The P300 speller, a communication device for
severely paralyzed patients, utilizes the event-related
potential function. Event-related desynchronization is
another function in which the power spectrum of the
EEG alpha band decreases following motor commands
]. Event-related desynchronization is often used in
rehabilitation systems. In the present study, we used
movement-related cortical potential (MRCP). MRCP
is the change in EEG signal resulting from the plan and
action of voluntary exercise [
]. MRCP is detectable
before and after exercise. In particular, MRCP that starts
about 800 ms before exercise is called the motor
readiness potential. We hypothesized that the motor
readiness potential would show the timing of when a motor
command occurs in the brain.
Support vector machine (SVM)
To detect MRCP, a pattern identification unit is required.
There are two kinds of pattern identification unit: One
uses a parametric method for which the probability
distribution of data is known in advance and another uses
a nonparametric method which requires collected data
because the probability distribution of data is unknown.
We used the nonparametric method because EEG data
are different for each patient.
We employed a support vector machine (SVM)
because it can divide known data into two classes [
]. Compared with other algorithms, SVM is suitable
to judge the two classes that the human tends to move
his body or not. Using this SVM system, we divided the
EEG data into data during rest and data during action.
The system needs to detect EEG data during action as a
SVM is a supervised learning method that can
construct pattern identification to two classes. SVM learns
the parameters required to maximize the margin from
where W is the weight parameter and h is the threshold.
If u > 0, sign(u) is 1. If u≤0, sign(u) is − 1. Figure 7 shows
the SVM classification. The Class 1 and Class 2 mean that
the label of each class is 1 and -1. The margin is the
distance between the separation plane and the closest data
point. SVM finds an optimal value of W to maximize the
Judgment of movement from acceleration
We recorded the time when the participant bent their
elbow using the accelerometer. We calculated the
Vx2 + Vy2 + Vz2 − 0.1
where Vxyz is the combined acceleration, and Vx, Vy and
Vz are the X-, Y- and Z-axis components of the
acceleration. We clarified the maximum value of the acceleration
during the rest period. We set the maximum acceleration
value during rest period as the threshold of the
movement starting judge. We judged when the acceleration
was over the threshold as the timing when the participant
bent the elbow.
To detect motor commands, we used EEG data from
0 to 2 s during each measurement session as the feature
quantity during rest. EEG data at 210 ms after the
acceleration of the wrist were considered the threshold for
motor command. There are two reasons to set this
threshold to 210 ms: One is the human cognitive delay. In this
experiment, the participant bent and their elbow based on
a signaling displayed on a monitor—the disappearance of
the red bar. Therefore, we considered that it would take
200 ms after the bar disappears for the participant to
recognize the bar disappearance. Another reason is machine
delay. There is a machine delay of 10 ms from motion
intention detection by the SVM to actuation of the FES.
From the above, we set the expected delay to 210 ms.
We clarified the identification ratio using
cross-validation of the data from 100 trials by each participant.
Figure 9 shows the cross-validation method. We divided
the original data into k blocks. Using the first block as
test data and the other data as training data, we
calculated the discrimination ratio. Next, using the second
data as test data and the other data as training data, we
calculated the discrimination ratio again. By repeating
the above procedure k times, we used the average of k
values for discrimination rate to estimate the accuracy of
Figure 10 shows the procedure for constructing the
SVM. The action time is the timing when the
acceleration of the wrist was over the maximum during rest.
The detection time is 210 ms before the action time. We
selected a value for k of 10, and the feature quantity was
all 14-ch EEG data during 210 ms from the detection
time to the action time.
To compare with other detection timings, we validate
the discrimination rate by each 70 ms via
cross-validation. The condition is 70, 140, 210, 280, 350 ms.
Results and discussion
Discrimination ratio by each detection timing is shown
in Table 2. This discrimination ratio is about 70%. The
results show that SVM could detect the MRCP
effectively. Particularly, the highest average is 72.9% at 280 ms
so we determine that the most appropriate detection
timing is 280 ms.
We clarified the discrimination ratio at 280 ms for each
participant, as shown in Fig. 11. The discrimination ratios
of participants A, B and C were 67.5, 69.4 and 81.9%,
respectively. The average was 72.9%; this is over the 45%
Using EEG data 210 ms after the red bar disappeared,
all discrimination data were at least 67%. This result was
above the 45% target value. This indicates that using EEG
data sorted by SVM, the proposed system can perform
FES on paralyzed patients with adjustable timing.
Repetitive facilitation exercise administered using the proposed
EEG system is potentially more successful than that
administered by a therapist.
In the present study, we carried out the experiment
by only three participants. We should conduct
experiments using more participants. In addition, the EEG
data were collected from only healthy subjects. For
some stroke patients, although the neural system is
different from the healthy subject, the sensory
recognition motor loops would be same as to healthy
because the neural system cannot feedback but can
feedforward. We should validate the detection using
EEG signals of paralyzed patients compared with the
healthy people. In future work, we will develop the
system using FES.
In the present study, we proposed a brain–machine
interface system to assist repetitive facilitation exercise. As a
result, the average accuracy was 72.9% and the highest at
the prediction time 280 ms. We conclude that 280 ms is
the most suitable to predict the judgment that a patient
intends to exercise or not. In future work, we will develop
this repetitive facilitation exercise assistance system.
SM conceived the study and drafted the manuscript. JT carried out all
experiments and analyzed the data. KY and MF participated in the research design
and sequence alignment. All authors read and approved the final manuscript.
The authors of this paper would like to thank Waseda University in Japan by
offering the funding support for this publication.
The authors declare that they have no competing interests.
Availability of data and materials
The datasets supporting the conclusions of this article are included within the
This research was supported in part by the Outstanding Graduate COE
Support Subsidy “Global Robot Academia (GRA)” from MEXT; in part by JSPS
KAKENHI under Grant Numbers JP16H07265, JP14J07226, 26242061; and in
part by the Council for Science, Technology and Innovation (CSTII),
Cross-ministerial Strategic Innovation Promotion Program (SIP) (funding agency: JST).
Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations
1. Estimation of the number of caregivers due to the number of stroke adults and stroke . Ministry of Health and Welfare Research Group , 2017 . Available: http://www.stroke-project. com/.
2. Embrey DG , Holtz SL , Alon G , Brandsma BA , McCoy SW . Functional electrical stimulation to dorsiflexors and plantar flexors during gait to improve walking in adults with chronic hemiplegia . Arch Phys Med Rehabil . 2010 ; 91 ( 5 ): 687 - 96 .
3. Langhorne P , Coupar F , Pollock A . Motor recovery after stroke: a systematic review . Lancet Neurol . 2009 ; 8 ( 8 ): 741 - 54 .
4. Barnes MP . Principles of neurological rehabilitation . J Neurol Neurosurg Psychiatry . 2003 ; 74 ( Suppl 4 ): iv3 - 7 .
5. Shindo K , Kawashima K , Ushiba J , Ota N , Ito M , Ota T , Kimura A , Liu M. Effects of neurofeedback training with an electroencephalogram-based brain-computer interface for hand paralysis in patients with chronic stroke: a preliminary case series study . J Rehabil Med . 2011 ; 43 ( 10 ): 951 - 7 .
6. Kawahira K , Shimodozono M , Ogata A , Tanaka N. Addition of intensive repetition of facilitation exercise to multidisciplinary rehabilitation promotes motor functional recovery of the hemiplegic lower limb . J Rehabil Med . 2004 ; 36 ( 4 ): 159 - 64 .
7. Shimodozono M , Noma T , Nomoto Y , Hisamatsu N , Kamada K , Miyata R , Matsumoto S , Ogata A , Etoh S , Basford JR , Kawahira K. Benefits of a repetitive facilitative exercise program for the upper paretic extremity after subacute stroke: a randomized controlled trial . Neurorehabil Neural Repair . 2013 ; 27 ( 4 ): 296 - 305 .
8. Bütefisch C , Hummelsheim H , Denzler P , Mauritz KH . Repetitive training of isolated movements improves the outcome of motor rehabilitation of the centrally paretic hand . J Neurol Sci . 1995 ; 130 ( 1 ): 59 - 68 .
9. Hoei T , Kawahira K , Fukuda H . Use of an arm weight-bearing combined with upper-limb reaching apparatus to facilitate motor paralysis recovery in an incomplete spinal cord injury patient: a single case report . J Phsyical Ther Sci . 2017 ; 29 : 176 - 80 .
10. Gheorghe L , Chavarriaga R , Millán JdelR. Steering timing prediction in a driving simulator task . In: Proceedings of annual international conference of the IEEE engineering in medicine and biology society. IEEE engineering in medicine and biology society; 2013 , p. 6913 - 16 .
11. Zhang H , Chavarriaga R , Gheorghe L , Millan JDR . Inferring driver's turning direction through detection of error related brain activity . In: Proceedings of annual international conference IEEE Engineering Medicine Biol Soc. EMBS; 2013 . p. 2196 - 99 .
12. Khaliliardali Z , Chavarriaga R , Andrei Gheorghe L , Millan JDR . Detection of anticipatory brain potentials during car driving . In: Proceedings of annual international conference IEEE engineering medicine biologly society EMBS ; 2012 . p. 3829 - 32 .
13. Choi K , Cichocki A . Control of a wheelchair by motor imagery in real time . In Intelligent data engineering and automated learning-IDEAL; 2008 . p. 330 - 7 .
14. Nittono H . Event-related potentials (ERPs) and cognitive activity : a new perspective from engineering psychology , p. 1 - 19 .
15. Pfurtscheller G , Lopes FH . Event-related EEG/MEG synchronization and desynchronization: basic principles . Clin Neurophysiol . 1999 ; 110 : 1842 - 57 .
16. Shibasaki H , Barrett G , Halliday E , Halliday AM . Components of the movement-related cortical potential and their scalp topography . Electroencephalogr Clin Neurophysiol . 1980 ; 49 ( 3-4 ): 213 - 26 .
17. Georgiev D , Lange F , Seer C , Kopp B , Jahanshahi M . Movement-related potentials in Parkinson's disease . Clin Neurophysiol . 2016 ; 127 ( 6 ): 2509 - 19 .
18. Suykens JAK , Vandewalle J . Least squares support vector machine classifiers . Neural Process Lett . 1999 ; 9 ( 3 ): 293 - 300 .
19. Tong S , Koller D. Support vector machine active learning with applications to text classification . J Mach Learn Res . 2001 ; 2 : 45 - 66 .
20. Furey TS , Cristianini N , Duffy N , Bednarski DW , Schummer M , Haussler D. Support vector machine classification and validation of cancer tissue samples using microarray expression data . Bioinformatics . 2000 ; 16 ( 10 ): 906 - 14 .
21. Etoh S. Evaluation of cortical plasticity with TMS . Japanese J Rehabil Med . 2002 ; 39 ( 12 ): 773 - 6 .